無人機擷取之多光譜影像應用於評估臺灣芻料燕麥營養價值的可行性

Abstract

芻料的營養價值 (Nutritive value, NV) 是一系列評估牧草品質的重要指標。然而，現階段測量農藝性狀以及營養價值的方法費時又費力，不利芻料作物育種的進行。透過穩健的統計方法，利用可反映與植物生理與型態有關之特定波長反射光譜之多光譜影像 (Multispectral imagery) 近似農藝性狀以及營養價值可為解決該問題的方法，但目前於臺灣尚未具備相關數據。因此，本研究目的為評估無人機擷取之多光譜影像應用於近似臺灣芻料燕麥營養價值的可行性。本實驗於2020年10月28日在臺灣大學附設農場以叢區種植四種具有不同性狀的燕麥品系（26號，43號，臺大選一號以及天鵝）。本試驗將田區劃分為十區，每個品系在區集內以三重複隨機種植四品系。當其中一個品系開始抽穗時，便於收穫當日利用無人機進行多光譜影像的拍攝，試驗期間總計進行八次拍攝，並使用反射光譜數據計算五十個不同的植生指數 (Vegetation indices, VIs) 作為預測變數以預測不同的性狀，並以循序向前特徵選取法 (Sequential forward floating selection, SFFS) 取得影響各性狀最重要的數個植生指數。結果顯示乾物質（Dry matter, DM）的預測模型表現最佳（R2 = 0.84）而其餘模型的表現則落在0.15至0.64之間。本研究提供從資料收集到模型建立的初步流程，未來透過更大量的數據、多年份的資料、光譜校正，與統計方法的調整等，期待更增加模型的預測能力。

Nutritive values (NV) are a series of important parameters for the evaluation of forage quality. However, current phenotyping of agronomic traits and NV for forage crops is labor-intensive and time-consuming. A potential strategy to overcome such tedious phenotyping is the application of multispectral imagery to approximate nutritive values with the help of robust statistical methods. Multispectral imagery contains information within specific wavelengths which are associated with biophysical parameters and plant morphology. However, there is no such information available for forage oat in Taiwan. Therefore, the objective of this study was to evaluate the possibility of approximating agronomic traits and NV of forage oat in Taiwan using multispectral imagery. On 28th October 2020, four oat lines (No. 26, No. 43, NTU Sel No 1. and Swan) with different characteristics were sown in 120 hill plots, allocated into ten blocks in the Experimental Farm of National Taiwan University. Within each block, four oat lines were allocated randomly and each line had three replicates. After either of the oat line within a block began heading, a total of eight times spectral measurements throughout the field trial were conducted with the help of an unmanned aerial vehicle on the day of harvest. In order to approximate each agronomic trait and NV, fifty vegetation indices (VIs) were derived from the spectral data to serve as predictor variables. Sequential forward floating selection was then utilized to choose the most important VIs for each trait. The result showed that the model for dry matter (DM) performed the best (R2 = 0.84) while the R2 for other traits ranged from 0.15 to 0.64. The present study laid a foundation workflow from data collection to model fitting, which can be further improved through multi-year data, spectral calibration, and improved statistical methods.